Thoughts on AI progress (Dec 2025)

Learning a standard tool and learning how a particular workplace operates are different requirements. In his December 2025 essay, Dwarkesh Patel argues that training models to use browsers or spreadsheets leaves a further challenge: acquiring the judgment and local knowledge that workers develop through feedback and experience.

He accepts that teaching common skills once can be efficient when many copies of a model reuse them. His objection concerns the changing, company-specific parts of a job that are difficult to prepare exhaustively in advance. This is his argument for treating learning on the job as a separate milestone when assessing progress toward broad automation.

“Models keep getting more impressive at the rate the short timelines people predict, but more useful at the rate the long timelines people predict.”

Dwarkesh Patel

“Human workers are valuable precisely because we don’t need to build schleppy training loops for every small part of their job.”

Dwarkesh Patel

“We’re losing money on every sale, but we’ll make it up in volume.”

Dwarkesh Patel

“we need something like a 1,000,000x scale-up of total RL compute to give a boost similar to a GPT level”

Toby Ord

“If these models were actually like humans on a server, they’d diffuse incredibly quickly.”

Dwarkesh Patel